Your agent already taught you. Now it writes the cards.

Set up the LearnReally MCP server for me.

WHAT IT IS: an open-source (MIT) MCP server published as the npm package @learnreally/mcp, run over stdio with npx. It adds three tools — create_deck, add_cards, list_decks — that write spaced-repetition cards to my LearnReally account.

1. THE KEY. The server authenticates with my personal API key, passed as the environment variable LEARNREALLY_API_KEY. Ask me for it and stop until I reply — never invent, guess, or reuse another key. If I don't have one, tell me to mint it at https://learnreally.com/settings (Account → API keys; it is shown once).

2. THE CONFIG. Work out which client you are running inside and register a stdio server named "learnreally" in that client's own format, passing my key as the LEARNREALLY_API_KEY environment variable (replace the placeholder in the sample). The shape most clients use:

{
  "mcpServers": {
    "learnreally": {
      "command": "npx",
      "args": ["-y", "@learnreally/mcp"],
      "env": {
        "LEARNREALLY_API_KEY": "lr_live_…your-key…",
        "LEARNREALLY_API_BASE_URL": "https://learnreally.com"
      }
    }
  }
}

The clients that differ:
- Claude Code — one command instead of a file:
claude mcp add --transport stdio learnreally \
  --env LEARNREALLY_API_KEY=lr_live_…your-key… \
  --env LEARNREALLY_API_BASE_URL=https://learnreally.com \
  -- npx -y @learnreally/mcp
- OpenAI Codex CLI — TOML at ~/.codex/config.toml: table [mcp_servers.learnreally] (snake_case) with a nested [mcp_servers.learnreally.env] table.
- Grok Build (xAI) — the same TOML, at ~/.grok/config.toml. Grok's in-app connectors take remote servers only, so a local one goes in the CLI.
- OpenCode — top-level "mcp", "type": "local", a single "command" array, and "environment" (not "env").
- OpenClaw — ~/.openclaw/openclaw.json, nested one level deeper: a "servers" map inside a top-level "mcp" object (not a root "mcpServers").
- Hermes (Nous Research) — YAML at ~/.hermes/config.yaml: top-level "mcp_servers", each server with command/args/env.
- VS Code (Copilot) — .vscode/mcp.json with the root key "servers" (not "mcpServers") and "type": "stdio".
- ChatGPT — not possible: it connects only to remote hosted MCP servers, never a local npx one.

3. VERIFY. Reload the client — usually I have to restart it myself, so tell me when — then call the list_decks tool. My decks (or "No decks yet") means it works. An authentication failure means the key; a startup failure means the command or LEARNREALLY_API_BASE_URL.

4. AFTER THAT. Tell me it is connected. From then on, whenever you teach me something I want to keep, call create_deck / add_cards and write active-recall cards for it — the server enforces the card rules and returns the reason for anything it rejects.